# Host paths.
MODEL_DIR=/mnt/llm_models/Kimi-K2.6-519B-NVFP4
LOG_DIR=/mnt/llm_models/logs/kimi-k2.6
METRICS_DIR=/mnt/llm_models/logs/kimi-k2.6/metrics

# Hugging Face source.
HF_MODEL_ID=0xSero/Kimi-K2.6-519B-NVFP4

# Container identity.
CONTAINER_NAME=sglang-kimi-k26
SGLANG_BASE_IMAGE=lmsysorg/sglang:v0.5.12.post1
SGLANG_IMAGE=local/kimi-k2-6-nvfp4-sglang:latest

# GPU placement. The verified host used four RTX PRO 6000 Blackwell GPUs.
CUDA_VISIBLE_DEVICES=0,2,3,4
CUDA_DEVICE_ORDER=PCI_BUS_ID

# Public API.
HOST=0.0.0.0
PORT=8000
SERVED_MODEL_NAME=kimi-k2.6

# SGLang model settings.
TP_SIZE=4
PP_SIZE=1
CONTEXT_LENGTH=262144
MAX_RUNNING_REQUESTS=4
MEM_FRACTION_STATIC=0.94
CHUNKED_PREFILL_SIZE=8192
MAX_PREFILL_TOKENS=16384
PAGE_SIZE=64
QUANTIZATION=modelopt_fp4
KV_CACHE_DTYPE=fp8_e4m3
ATTENTION_BACKEND=flashinfer
MOE_RUNNER_BACKEND=auto
TOOL_CALL_PARSER=kimi_k2
REASONING_PARSER=kimi_k2
CHAT_TEMPLATE=/workspace/model/chat_template.jinja
CUDA_GRAPH_MAX_BS=4
WATCHDOG_TIMEOUT=3600
MODEL_LOADER_THREADS=32

# Runtime features.
ENABLE_MULTIMODAL=1
TRUST_REMOTE_CODE=1
PRE_WARM_NCCL=1
ENABLE_METRICS=1
LOG_REQUESTS=1
LOG_REQUESTS_LEVEL=3

# Docker runtime.
SHM_SIZE=96g

# Compatibility env from the verified launch.
NCCL_DEBUG=WARN
PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True
TORCH_CUDA_ARCH_LIST=12.0
NCCL_P2P_LEVEL=PIX
NCCL_CUMEM_HOST_ENABLE=0
OMP_NUM_THREADS=8
NCCL_NVLS_ENABLE=0
NCCL_SOCKET_IFNAME=lo
GLOO_SOCKET_IFNAME=lo
SGLANG_ENABLE_DEEP_GEMM=0
SGLANG_DISABLE_DEEP_GEMM=1
SGLANG_ENABLE_JIT_DEEPGEMM=0
SGLANG_SET_CPU_AFFINITY=1
FLASHINFER_DISABLE_VERSION_CHECK=1
TORCH_NCCL_BLOCKING_WAIT=1
TORCH_NCCL_ASYNC_ERROR_HANDLING=1
SGLANG_MOE_NVFP4_DISPATCH=1
SAFETENSORS_FAST_GPU=1
NCCL_P2P_DISABLE=0
NCCL_IB_DISABLE=1